How to use totalcontrastiveVI?

[HelloWorldLTY]

Hi, do we have tutorials for totalcontastiveVI? Thanks.

[ethanweinberger]

Hi! Currently we don't have a publicly available tutorial, but I can create a notebook based on the ECCITE-seq analysis from our paper- I should have the notebook ready by sometime next week.

[HelloWorldLTY]

Thanks, I have successfully reproduced your work for running totalcontrastiveVI and TotalVI, based on your contrastive-vi reproduce folder.

I actually further have a question about running totalvi:

https://github.com/suinleelab/contrastiveVI-reproducibility/blob/43efaab1990410179f0005ba8c6dd411cc8789a4/scripts/run_experiment.py#L356C1-L361C14

It seems that you did not include any observation variables (like replicate as a key in batch) for running totalvi, and I guess it is th reason why totalvi is confounded by replicates or cell cycles. Do you think it is a reasonable inference? Thanks.

[ethanweinberger]

Great question! We actually explored this question in depth in Supplementary Note 9 from the contrastiveVI manuscript (https://static-content.springer.com/esm/art%3A10.1038%2Fs41592-023-01955-3/MediaObjects/41592_2023_1955_MOESM1_ESM.pdf).

In short: we found that including these additional covariates did somewhat mitigate the issues with running totalVI naively. However, there was still increased confounding by cell cycle/replicate/etc. compared to totalContrastiveVI, and separation of the perturbation groups was worse than with totalContrastiveVI (see e.g. the IRF1 cells in Supplementary Figure 31). Also, this workflow (i.e., providing totalVI with covariates), assumes that all potential confounders are known a priori, while totalContrastiveVI doesn't have this requirement.